Neural models are required before synthesis can begin. This will download approximately 300 MB of model weights from the HF Hub and cache them locally.
SpeechT5 TTS + HiFi-GAN Vocoder~300 MB
Voice Profile
Voice cloning is not available in-browser.
SpeechT5 requires 512-dimensional x-vector speaker embeddings produced by a dedicated speaker-recognition model (e.g., SpeechBrain ECAPA-TDNN). These cannot be extracted client-side from an audio file using general-purpose wav2vec2 feature extractors — the embedding spaces are incompatible.
True cloning requires server-side x-vector extraction or a pre-computed embedding. The two presets above are official CMU ARCTIC x-vectors cached from the HF Hub. We have created a contained enviornment @ inky9.lovable.app which is now available as a fully functioning alternative.
PROCESS: Client-side synthesis using Xenova/speecht5_tts (16 kHz) via @huggingface/transformers. Speaker embeddings are pre-computed 512-dim x-vectors from the CMU ARCTIC dataset.
VOICE CLONING: Not supported client-side. Requires a dedicated x-vector/ECAPA speaker encoder whose output distribution matches SpeechT5's training data. General-purpose wav2vec2 feature extractors produce incompatible embeddings.
access https://inky9.lovable.app/ for more options
All operations remain within the local browser nexus. No data is transmitted to external servers.